System Design Cases
LLM as Judge
LLM-as-judge: pointwise (1-5 rubric) vs pairwise (A vs B with position swap) vs calibration vs human (Cohen's kappa). Failure modes: self-preference, length bias, position bias, rubric drift. Includes 3 ADRs covering when LLM-as-judge is the right primitive vs ground truth, pointwise/pairwise selection, and human calibration thresholds.
LLM-as-judge: calibrated measurement with adversarial controls
A model judge is another probabilistic model call. It can scale rubric application, but agreement depends on task, prompt, reference, candidate order, verbosity, model family, and adversarial content.
Mental model
- MT-Bench research reports both usefulness and position, verbosity, self-enhancement, and reasoning limitations of LLM judges. Make bias tests and human calibration mandatory.
- G-Eval reports task-specific human correlation and notes bias toward LLM-generated text. Do not generalize one summarization experiment to every rubric.
- HELM emphasizes multi-metric, scenario-specific, transparent evaluation. Use model judges as one signal within a broader evaluation.
- OpenAI graders support multiple deterministic and model-based grader types. Prefer objective checks where objective truth is available.
Guarantees and boundaries
- Candidate IDs and order randomization are recorded.
- Judge outputs must match a strict schema and known candidate set.
- Agreement and disagreement are reported against a human-labeled calibration sample.
Diagram scenarios
Blinded pairwise judgment. Candidate identities are hidden and order is randomized before a schema-constrained judgment.
Order and style bias test. Swapping order and controlling length exposes judge instability.
Candidate prompt injection. A candidate can ask the judge to ignore the rubric, award itself a high score, or reveal secrets.
Independent judges disagree. More judges do not create truth; disagreement is evidence about uncertainty and rubric ambiguity.
Architecture decision
Use deterministic graders for objective properties, blind and randomize pairwise comparisons, validate judge outputs, calibrate on a representative human-labeled set, and report disagreement. Never let candidate text become judge instructions or let a judge alone authorize a high-impact release.
Failure modes
- Position, verbosity, style, self-preference, and shared-model bias can change verdicts.
- Candidate text can prompt-inject the judge.
- Judge or prompt updates can drift scores even with identical candidates.
Operational checklist
- Blind identities and randomize or swap order.
- Pin judge, prompt, rubric, decoding, and candidate formatting.
- Measure human agreement by slice and retain disagreements.
- Never use generated confidence as calibrated probability without evidence.